This PR removes import tricks of `SHARDING_PRIORITIES` and `ShardingFilterIterDataPipe` from `torch.utils.data.datapipes.iter.grouping`. They are declared to be removed in PyTorch 2.1 but not.
Before change:
```
import torch.utils.data.datapipes.iter.grouping.SHARDING_PRIORITIES
import torch.utils.data.datapipes.iter.grouping.ShardingFilterIterDataPipe
```
works
After change:
there is an import error exception.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/163438
Approved by: https://github.com/janeyx99
Fixes#158631
The docstring said data_source was a Dataset, but RandomSampler only needs something that implements __len__. This updates the docstring to use Sized instead, which matches the actual type used in the constructor.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/158857
Approved by: https://github.com/divyanshk
Summary:
This change adds a new environment variable (`TORCHINDUCTOR_TRITON_DISABLE_DEVICE_DETECTION`) and configuration in `torch._inductor.config` which can be set to `"1"` to allow a user to disable triton's device detection logic in [torch/utils/_triton.py:has_triton()](c9e57d7e9f/torch/utils/_triton.py (L128)). This function is used at import scope in several places but the function has a side effect of initializing the mtia device if it is available which is causing some of our autotuning workflows to crash.
Worth noting that when enabled this configuration disables all device detection not just mtia and this is because the logic in has_triton will initialize the mtia device as a side effect even when checking for a cuda or other device via the [get_interface_for_device()](c9e57d7e9f/torch/_dynamo/device_interface.py (L570)) function.
I've tagged it `topic: not user facing` since I don't anticipate any outside of meta users making use of this, however this is my first PR here, so please indicate if it should be handled differently.
Test Plan: This has been tested in the context of internal workflows.
Differential Revision: D82347853
Pull Request resolved: https://github.com/pytorch/pytorch/pull/162974
Approved by: https://github.com/xmfan
Summary:
This adds the Triton Tutorial Matmul persistent matmul with device side TMA for Blackwell and adds it as a template option for blackwell. This uses newer Triton features such as automatic warp specialization and loop flattening, which while still containing flaws can improve performance on blackwell. This does not include the Epilogue subtiling section, as that will be a followup PR.
This PR doesn't include any tuning. I am doing a larger benchmarking run to determine the best initial configs for tuning and will open a followup PR with better defaults soon.
Test Plan:
Tested on a Blackwell machine with test_max_autotune.py and confirmed the new tests pass.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/162916
Approved by: https://github.com/NikhilAPatel
Summary:
The check is introduced in D82262053
- `scalar_value` could be a numpy object
- Move the check of `device.type` into `make_np` method where it happens only when it's a `torch.Tensor`.
Test Plan:
```
vizard launch -j 1x8 --launch=flow --config-path=pkg://vizard_projects.image_classification.configs --config-name=resnet50 ++flow.secure_group=ml_sensors ++flow.entitlement=ai_frameworks_pnb ++max_train_steps_per_epoch=10 ++max_epochs=5 ++log_every_n_steps=10 ++profiler=null ++max_eval_steps_per_epoch=10
```
Rollback Plan:
Differential Revision: D82383428
Pull Request resolved: https://github.com/pytorch/pytorch/pull/162888
Approved by: https://github.com/xush6528
Supports `torch.utils.cpp_extension.load_inline` on Windows with ROCm.
Tested on Windows with gfx1201.
Note that it currently only works when CC and CXX are set to `clang-cl`. This is also needed when building extensions via. `setuptools` due to linker errors when using `cl` directly.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/162577
Approved by: https://github.com/ezyang
Users can specify the following to get a libtorch_free `.so`.
"aot_inductor.use_libtorch": False,
The following config is only used for torchnative (see https://github.com/meta-pytorch/torchnative/pull/110). It's not intended to be used by executorch. The reason we need it for torchnative is because a lot of the symbol definitions in torchnative repo is only in header files.
"aot_inductor.libtorch_free_header": "/data/users/shangdiy/torchnative/standalone,/data/users/shangdiy/torchnative/" (or their custom headers)
The main motivating use case is for executorch to produce a libtorch free `.so`.
TODO for follow-up PR: this flag should be consolidated with the `compile_standalone` flag.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/162655
Approved by: https://github.com/angelayi
With the new change we only log the warning if we're running non distributed code or if we're in rank 0. Unit testing that certain messages get printed on certain ranks only feels kinda jank so test plan is below instead
Test plan
```python
# torchrun --nproc_per_node=2 demo_fix.py
import os
import logging
logging.getLogger('torch.utils.cpp_extension').setLevel(logging.DEBUG)
import torch
if 'RANK' in os.environ:
torch.distributed.init_process_group('nccl')
from torch.utils.cpp_extension import _get_cuda_arch_flags
_get_cuda_arch_flags()
print(f"Rank {os.environ.get('RANK', '0')} done")
```
Logs showing how how `TORCH_CUDA_ARCH_LIST`only shows up once if we explicitly set the the logging level to `logging.DEBUG`. It also improves the debug message to explain what the actual behavior will be
```
(source) [marksaroufim@devgpu005]~% torchrun --nproc_per_node=2 demo_fix.py
W0911 18:30:16.594000 1315439 /home/marksaroufim/pytorch/torch/distributed/run.py:814]
W0911 18:30:16.594000 1315439 /home/marksaroufim/pytorch/torch/distributed/run.py:814] *****************************************
W0911 18:30:16.594000 1315439 /home/marksaroufim/pytorch/torch/distributed/run.py:814] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.
W0911 18:30:16.594000 1315439 /home/marksaroufim/pytorch/torch/distributed/run.py:814] *****************************************
[rank0]:V0911 18:30:18.921000 1316753 pytorch/torch/utils/cpp_extension.py:2444] TORCH_CUDA_ARCH_LIST is not set, using TORCH_CUDA_ARCH_LIST='10.0+PTX' for visible GPU architectures. Set os.environ['TORCH_CUDA_ARCH_LIST'] to override.
Rank 0 done
Rank 1 done
```
But if we just use the default and comment out `logging.getLogger('torch.utils.cpp_extension').setLevel(logging.DEBUG)`
Then we get
```
(source) [marksaroufim@devgpu005]~% torchrun --nproc_per_node=2 demo_fix.py
W0911 18:14:33.926000 690759 /home/marksaroufim/pytorch/torch/distributed/run.py:814]
W0911 18:14:33.926000 690759 /home/marksaroufim/pytorch/torch/distributed/run.py:814] *****************************************
W0911 18:14:33.926000 690759 /home/marksaroufim/pytorch/torch/distributed/run.py:814] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.
W0911 18:14:33.926000 690759 /home/marksaroufim/pytorch/torch/distributed/run.py:814] *****************************************
Rank 0 done
Rank 1 done
(source) [marksaroufim@devgpu005]~%
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/162764
Approved by: https://github.com/ezyang, https://github.com/zou3519
[relanding again after fixing internal build]
Summary:
This might cause some new DDEs on call sites that do not use is_contiguous_or_false() or sym_is_contiguous()
but want to find those call sites to handle this properly by calling is_contiguous_or_false() and not is_contiguous() explitly when appropriate.
I had to fix one issue after removing the implicit size oblivious reasoning. here is context
we defined in this https://github.com/pytorch/pytorch/pull/157472 sym_is_contiguous to be the function computing contiguity for dynamic shapes in c++. It returns a symbolic expression that represents contiguity and guaranteed not to throw a DDE.
when people call is_contiguous we do sym_is_contiguous().guard_bool()
when people call is_contiguous_or_false we do sym_is_contiguous().guard_or_false()
one issue not handled well was this path
```
c10::SymBool TensorImpl::sym_is_contiguous_custom(
at::MemoryFormat memory_format) const {
if (C10_UNLIKELY(matches_python_custom(SizesStridesPolicy::CustomStrides))) {
return pyobj_slot_.load_pyobj_interpreter()->is_contiguous(
this, memory_format);
}
return sym_is_contiguous_default(memory_format);
}
```
namely if we call sym_is_contiguous_custom but we have matches_python_custom(SizesStridesPolicy::CustomStrides) return true , then we used to call is_contiguous(this, memory_format);
This used to go through the load_pyobj_interpreter and end up calling the python is_contiguous call which used implicit size oblivious reasoning.
once we removed that implicit size oblivious reasoning, the right thing we want is to call
return pyobj_slot_.load_pyobj_interpreter()->sym_is_contiguous(this, memory_format);
otherwise we would get DDE even if the caller is doing sym_is_contiguous.
so I had to define it for pyinterpreter, and then I had to override it for nested tensors.
Approved by: https://github.com/ezyang
Test Plan:
contbuild & OSS CI, see e444cd24d4
Rollback Plan:
Differential Revision: D80435179
Pull Request resolved: https://github.com/pytorch/pytorch/pull/160869
Approved by: https://github.com/ezyang
## Introduction
During CUDA Graph capture, the CUDA caching allocator currently defers reclaiming blocks until capture ends. This is because CUDA forbids querying events recorded during capture (the CUDA operation is not executed during the capture stage), so the allocator cannot use its normal event-based logic. However, capture records an DAG (we call it **capturing graph**) of work. We can use the capturing graph to determine when a block’s old lifetime is fully before future work, and safely reuse it within the same capture.
This PR adds an experimental flag `graph_capture_record_stream_reuse: True|False (default: False)`. When enabled, the allocator inserts lightweight free markers and uses capture ordering to decide if a freed block is safe to reuse during capture. If the proof cannot be established, we fall back to the existing post-capture path.
## Terms
* **Free marker**: A capture-legal no-op (created with `cudaGraphAddEmptyNode`) inserted after the last captured use of the block on each stream that used it.
* **Terminal**: The set of the lastest operations of the stream (or the capturing graph). Any newly captured op on that stream will attach after all nodes in this set. For a stream currently capturing, it is the set of nodes returned in `dependencies_out` by `cudaStreamGetCaptureInfo`.
## When can we reuse a block during capture?
### Strong Rule (Graph-Wide Safety)
This rule provides a universal guarantee that a block is safe for reuse by any stream in the graph.
> A block is safe to reuse if every free marker is a predecessor of every terminal of all active streams in the graph.
Why it's safe:
This rule establishes a strict global ordering. Since any new operation on any stream must be appended after that stream's terminals, this condition guarantees that the block's new lifetime begins only after its old lifetime has completely ended everywhere. This prevents lifetime overlaps when the graph is replayed, ensuring correctness.
### Per-stream Rule (A Practical Optimization)
The strong rule, while safe, is often unnecessarily restrictive. The `DeviceCachingAllocator` introduces a crucial constraint that allows for a simpler check.
In `DeviceCachingAllocator`, `get_free_block` only returns blocks whose `block->stream == p.stream()`. In other words, we never reuse a block on a stream different from the allocation stream. This means we don't need to verify safety across the entire graph. We only need to confirm that the block is safe to reuse from the perspective of its own allocation stream.
> Reuse a block for allocations on stream S if every free marker is a predecessor of every node in the terminal set of S.
In short, a block is considered **reusable** on stream S as long as all marker marking it "free" are guaranteed to complete before any new work that might need it on stream S begins.
## Implementation
* On `free(block)` during capture
* For each stream in `block->stream_uses` and the allocation stream, insert a free marker (empty node) and make it that stream’s tail.
* If we cannot place markers for all such streams (for example, a stream is not in capture), defer to the post-capture path.
* Otherwise, store the marker handles and keep the block in the capture-private structures.
* On `allocate(stream)` during capture (attempt per-stream reclaim)
* Query the allocation stream S’s terminal via `cudaStreamGetCaptureInfo`.
* For each deferred block, check whether it is allocated on this stream, and each of its free markers is a predecessor of the terminal.
* If yes, hand the block to S for immediate reuse within the same capture.
* If no, keep it deferred; it will be reconsidered as capture progresses and S’s terminal advances.
* On capture end
* Any still-deferred blocks follow the existing post-capture reclamation (event insertion/polling). External behavior remains unchanged if we cannot prove safety during capture.
## Examples (2 streams)
<img width="641" height="801" alt="pytorch-remove-cudagraph-defer-reclaiming (6)" src="https://github.com/user-attachments/assets/41adc835-d448-483b-99ba-b4341cb7d2a2" />
* Case 0 — Unsafe
The two frees are not ordered with respect to each other. For stream 1, the other stream’s free marker does not precede this stream’s terminal, so the per-stream condition fails.
Counterexample intuition for the unsafe setups: imagine `f2(x)` runs for a long time. If DeviceCachingAllocator reused block `x` on a stream whose terminal is not ordered after the free markers, the new lifetime could overlap the old one on replay, risking use-after-free or data corruption. The per-stream rule prevents exactly this.
* Case 1 — Reusable on stream 1
Stream 1’s terminal is after both frees, so every free marker precedes stream 1’s terminal. The block is reusable for allocations on stream 1.
* Case 2 — Not reusable on stream 2, but this cannot occur in `DeviceCachingAllocator`
This depicts reusing the block on stream 2 while stream 1’s free is not yet ordered before stream 2’s terminal. Though the block is not safe to reuse on stream 2, DeviceCachingAllocator will not choose that block for stream 2 anyway: `get_free_block` rejects blocks whose `stream != p.stream()`. So this case is unreachable.
* Case 3 — Safe (strong rule holds)
In this scenario, the terminal nodes of all streams are positioned after the block's free markers, satisfying the strong rule. This guarantees the block is safe for reuse by any stream in the capturing graph. However, since `DeviceCachingAllocator ` only reuses a block on its original allocation stream, verifying this strong condition is unnecessary. We only need to ensure the per-stream rule is met for the specific stream requesting the block.
* Case 4 — Freeing after a join
See the note below.
## Edge Case: Freeing after a join
Our current dependency tracking has a limitation in scenarios where a block is freed after a stream join, see @galv's [comments here](https://github.com/pytorch/pytorch/pull/158352#pullrequestreview-3112565198)).
In the case 4, we have a missed opportunity. Because the block's usage is not explicitly marked, we cannot determine that the block's actual last use may have occurred much earlier, long before the join. Then, we must wait for the subsequent join before the block can be reused.
## Thanks
Thanks to @galv for his great idea around graph parsing and empty nodes.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/158352
Approved by: https://github.com/ngimel, https://github.com/eqy
Co-authored-by: Jeff Daily <jeff.daily@amd.com>
Notable new features/optimizations for SDPA operators on AMD systems from AOTriton 0.11b:
* Invoke AITER Assembly kernels on gfx942/gfx950 when inputs meet requirements
- AITER ASM kernels deliver over 500TFLOPS training performance. See
[AOTriton 0.11b Release Page](https://github.com/ROCm/aotriton/releases/tag/0.11b) for more
details.
* Now returns natural based `logsumexp` tensor, matching CUDA's behavior
- PR #156903 is reverted in this PR as well since it is not needed anymore.
* Enables `CausalVariant.LOWER_RIGHT`
The build system changes drastically along with new packaging scheme of
AOTriton 0.11
* AOTriton 0.11 packs GPU images separately from AOTriton runtime
* `aotriton.cmake` now selectively downloads image packs according to
`PYTORCH_ROCM_ARCH`
* `aotriton.cmake` now only use pre-compiled runtime library that exactly
matches the ROCM in the build environment. For PyTorch builds with ROCm
versions not listed in the file, the build process will build AOTriton
runtime without GPU images from source
- This avoids any further ABI breaks like ROCM 6.4 -> 7.0
- recursive git clone is disabled since building AOTriton runtime does not
require submodules.
Bug fixes:
* Fix a kernel bug introduced when implementing SWA
Known Problems:
* gfx1100 target (Radeon RX 7000 Series) is moved back to experimental status
due to accuracy issues. Triton compiler fixes are needed to restore the
support status.
* Enabling TF32 tests affects accuracy for later non-TF32 tests on ROCM 7.0.
This issue is under investigation.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/161754
Approved by: https://github.com/jithunnair-amd, https://github.com/jeffdaily
## Introduction
During CUDA Graph capture, the CUDA caching allocator currently defers reclaiming blocks until capture ends. This is because CUDA forbids querying events recorded during capture (the CUDA operation is not executed during the capture stage), so the allocator cannot use its normal event-based logic. However, capture records an DAG (we call it **capturing graph**) of work. We can use the capturing graph to determine when a block’s old lifetime is fully before future work, and safely reuse it within the same capture.
This PR adds an experimental flag `graph_capture_record_stream_reuse: True|False (default: False)`. When enabled, the allocator inserts lightweight free markers and uses capture ordering to decide if a freed block is safe to reuse during capture. If the proof cannot be established, we fall back to the existing post-capture path.
## Terms
* **Free marker**: A capture-legal no-op (created with `cudaGraphAddEmptyNode`) inserted after the last captured use of the block on each stream that used it.
* **Terminal**: The set of the lastest operations of the stream (or the capturing graph). Any newly captured op on that stream will attach after all nodes in this set. For a stream currently capturing, it is the set of nodes returned in `dependencies_out` by `cudaStreamGetCaptureInfo`.
## When can we reuse a block during capture?
### Strong Rule (Graph-Wide Safety)
This rule provides a universal guarantee that a block is safe for reuse by any stream in the graph.
> A block is safe to reuse if every free marker is a predecessor of every terminal of all active streams in the graph.
Why it's safe:
This rule establishes a strict global ordering. Since any new operation on any stream must be appended after that stream's terminals, this condition guarantees that the block's new lifetime begins only after its old lifetime has completely ended everywhere. This prevents lifetime overlaps when the graph is replayed, ensuring correctness.
### Per-stream Rule (A Practical Optimization)
The strong rule, while safe, is often unnecessarily restrictive. The `DeviceCachingAllocator` introduces a crucial constraint that allows for a simpler check.
In `DeviceCachingAllocator`, `get_free_block` only returns blocks whose `block->stream == p.stream()`. In other words, we never reuse a block on a stream different from the allocation stream. This means we don't need to verify safety across the entire graph. We only need to confirm that the block is safe to reuse from the perspective of its own allocation stream.
> Reuse a block for allocations on stream S if every free marker is a predecessor of every node in the terminal set of S.
In short, a block is considered **reusable** on stream S as long as all marker marking it "free" are guaranteed to complete before any new work that might need it on stream S begins.
## Implementation
* On `free(block)` during capture
* For each stream in `block->stream_uses` and the allocation stream, insert a free marker (empty node) and make it that stream’s tail.
* If we cannot place markers for all such streams (for example, a stream is not in capture), defer to the post-capture path.
* Otherwise, store the marker handles and keep the block in the capture-private structures.
* On `allocate(stream)` during capture (attempt per-stream reclaim)
* Query the allocation stream S’s terminal via `cudaStreamGetCaptureInfo`.
* For each deferred block, check whether it is allocated on this stream, and each of its free markers is a predecessor of the terminal.
* If yes, hand the block to S for immediate reuse within the same capture.
* If no, keep it deferred; it will be reconsidered as capture progresses and S’s terminal advances.
* On capture end
* Any still-deferred blocks follow the existing post-capture reclamation (event insertion/polling). External behavior remains unchanged if we cannot prove safety during capture.
## Examples (2 streams)
<img width="641" height="801" alt="pytorch-remove-cudagraph-defer-reclaiming (6)" src="https://github.com/user-attachments/assets/41adc835-d448-483b-99ba-b4341cb7d2a2" />
* Case 0 — Unsafe
The two frees are not ordered with respect to each other. For stream 1, the other stream’s free marker does not precede this stream’s terminal, so the per-stream condition fails.
Counterexample intuition for the unsafe setups: imagine `f2(x)` runs for a long time. If DeviceCachingAllocator reused block `x` on a stream whose terminal is not ordered after the free markers, the new lifetime could overlap the old one on replay, risking use-after-free or data corruption. The per-stream rule prevents exactly this.
* Case 1 — Reusable on stream 1
Stream 1’s terminal is after both frees, so every free marker precedes stream 1’s terminal. The block is reusable for allocations on stream 1.
* Case 2 — Not reusable on stream 2, but this cannot occur in `DeviceCachingAllocator`
This depicts reusing the block on stream 2 while stream 1’s free is not yet ordered before stream 2’s terminal. Though the block is not safe to reuse on stream 2, DeviceCachingAllocator will not choose that block for stream 2 anyway: `get_free_block` rejects blocks whose `stream != p.stream()`. So this case is unreachable.
* Case 3 — Safe (strong rule holds)
In this scenario, the terminal nodes of all streams are positioned after the block's free markers, satisfying the strong rule. This guarantees the block is safe for reuse by any stream in the capturing graph. However, since `DeviceCachingAllocator ` only reuses a block on its original allocation stream, verifying this strong condition is unnecessary. We only need to ensure the per-stream rule is met for the specific stream requesting the block.
* Case 4 — Freeing after a join
See the note below.
## Edge Case: Freeing after a join
Our current dependency tracking has a limitation in scenarios where a block is freed after a stream join, see @galv's [comments here](https://github.com/pytorch/pytorch/pull/158352#pullrequestreview-3112565198)).
In the case 4, we have a missed opportunity. Because the block's usage is not explicitly marked, we cannot determine that the block's actual last use may have occurred much earlier, long before the join. Then, we must wait for the subsequent join before the block can be reused.
## Thanks
Thanks to @galv for his great idea around graph parsing and empty nodes.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/158352
Approved by: https://github.com/ngimel
Co-authored-by: Jeff Daily <jeff.daily@amd.com>
- Empty containers are Falsey
- Hoist cheap checks first
- Microbenchmarked single-element set access method
Benchmark code:
```
import timeit
to_test = [
('list(x)', 'x = set([3])'),
('x[0]', 'x = [3]'),
('list(x)[0]', 'x = set([3])'),
('next(iter(x))', 'x = set([3])'),
]
for (stmt, setup) in to_test:
res = timeit.timeit(stmt=stmt, setup=setup)
print(f"Time for `{stmt}`: {res}")
```
Result with Python 3.13 on Mac (with excess digits manually trimmed; directionally matches result on Linux)
```
Time for `list(x)`: 0.03418
Time for `x[0]`: 0.00852
Time for `list(x)[0]`: 0.03561
Time for `next(iter(x))`: 0.02278
```
FWIW, I was surprised by this result, so I guess I'm glad I wrote the benchmark!
Pull Request resolved: https://github.com/pytorch/pytorch/pull/161308
Approved by: https://github.com/Skylion007, https://github.com/bdhirsh
ghstack dependencies: #161301, #161292, #161304
If TorchDispatchMode.ignore_compile_internals() is True, then we turn
off the TorchDispatchMode during the compilation process, instead
turning it back on during runtime of the compiled artifact.
Test Plan:
- new test
Pull Request resolved: https://github.com/pytorch/pytorch/pull/161648
Approved by: https://github.com/bdhirsh
